In model-based testing the size of the used model has a great impact on the time for computing test cases. In model checking, dependence relations have been used in slicing of specifications in order to obtain reduced models pertinent to criteria of interest. In specifications described using state based formalisms slicing involves the removal of transitions and merging of states thus obtaining a structural modified specification. Using such a specification for model based test case generation where sequences of transitions represent test cases might provide traces that are not valid on a correctly behaving implementation. In order to avoid such trouble, we suggest the use of control, data and communication dependences for identifying parts of the model that can be excluded so that the remaining specification can be safely employed for test case generation. This information is included in test purposes which are then used in the test case generation process. We present also first empirical results obtained by using several models from industry and literature.